Improved visualization of perivascular spaces on T2-weighted imaging with deep learning-based denoising and...

This prospective study evaluated the ability of a deep learning-based denoising followed by super-resolution function (SR-DL) to improve the visualization of perivascular spaces (PVSs) on T2-weighted imaging (T2WI). Ten healthy volunteers underwent brain MRI using T2WI with three acquisition voxel sizes (1.0 × 1.0 mm, 0.8 × 0.8 mm, and 0.5 × 0.5 mm), each reconstructed to half-size in-plane resolution. For each acquisition, image sets were generated using compressed-sensing sensitivity-encoding (CS-SENSE) and SR-DL, and a fully sampled high-resolution dataset with a reduction factor of 1 was obtained as a reference. Quantitative evaluation consisted of counting PVSs in the centrum semiovale and basal ganglia by two radiologists. In addition, contrast ratio (CR) between PVSs and adjacent brain parenchyma was measured using region-of-interest analysis. The utility of SR-DL T2WIs for PVS segmentation was also assessed using an existing automated method on both CS-SENSE and DL-SR images. Qualitative visibility was evaluated on a four-point scale. SR-DL significantly increased the number of detectable PVSs compared with CS-SENSE at medium- and high-resolution voxel sizes for both readers in both regions (P < 0.05 to < 0.001). SR-DL demonstrated significantly higher CRs between PVSs and adjacent white matter than CS-SENSE across all assessments (all P < 0.05). Automated PVS segmentation also showed significantly more detected PVSs with SR-DL than with CS-SENSE at all voxel sizes (all P < 0.05). Qualitative assessment also demonstrated consistently higher visibility scores with SR-DL across all voxel sizes (P < 0.05). These findings indicate that SR-DL reconstruction markedly improves both quantitative detectability and qualitative visibility of PVSs on T2WI without prolonging acquisition time, potentially facilitating more reliable PVS assessment and enhancing the utility of PVS-based imaging biomarkers.